Probabilistic Tracking with Database Systems

Probabilistic tracking is a standard technique for estimating the state of moving objects from sensor data. With increasing numbers of sensors, however, sensor data has become overwhelming. This has led to the usage of database systems for sensor data processing. Tracking tasks so far have been accomplished in external components. In this thesis, a solution for integrating tracking methods into generic relational database systems (RDBMS) is provided. For four classes of representative sensor data processing algorithms, relational and deductive rules are derived for implementing probabilistic tracking algorithms into RDBMS. These rules are declarative, descriptive formulations for sensor data processing. For necessary additional functionalities concerning probabilistic functions and matrix implementations, several implementation solutions are discussed and evaluated. A new phase concept on data streams enhances the implementation of anomaly detection. An analysis of the asymptotical runtime behavior of the deductive implementation shows its superlinear runtime which is unwanted for real-time applications. As a solution, incremental update propagation methods are applied which restore the linear runtime behavior. The methods are tested on a prototypical system for the detection of critical situations in airspace monitoring. The test proves the suitability for real-world scenarios, in this case allowing the processing of flight data for all German airspace in real time on standard computer hardware.